Exploring individual, social and environmental factors related to physical activity: a network analysis

Participants

Participants (N=20 611 Japanese speaking adults) were recruited from a sample-pool database; more than a million online panels had been registered to this database. The sample size was determined for practical and pragmatic reasons: (1) we expected that a large sample size would be required to estimate a network of 45 variables (PA and 44 correlates, with 990 possible associations), and therefore, we aimed for the maximum affordable number considering the financial and human resources that we had; and (2) the collected data were analysed for other purposes (not reported here), which specifically focused on a particular group of participants (eg, mHealth app users) and required a sufficiently large size for the subsample. Eligible participants (ie, being aged >18 years, having a good command over Japanese and residency in Japan) received an invitation for two online surveys separated by a month (early 2023). In the first survey, participants completed questionnaires regarding demographics, levels and readiness for PA, psychological characteristics, social support and environmental factors relevant for PA. The second survey encompassed the current health status and medical histories. For each survey, participants received a small compensation for their participation (online shopping voucher). Of the 20 611 participants, 19 039 completed both the surveys. Data of 34 participants were deemed unreliable; that is, those reported: (1) a height of ≤100 cm; (2) a weight of ≤10 kg; and (3) total active time of >24 hours per day. The remaining 19 005 responses were submitted for statistical analyses. All participants provided informed consent. We reported the results in accordance with Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) statement16 and the reporting standards for psychological network analyses in cross-sectional data.17

Equity, diversity and inclusion statement

Our team consisted of four Japanese men: one junior researcher and three senior researchers. The study population covered a wide range of age groups with balanced gender. However, the online survey was conducted in Japan exclusively and written in Japanese. We assumed that participants had good language command and internet literacy, which may have affected the demographics of participants and thus may limit the generalisability of the results (eg, individuals with lower socioeconomic status or from more marginalised communities may not be included).

Measures

Overall, we based variable selection on systematic literature reviews, such as Trost et al,12 which summarised and categorised PA-relevant factors into demographic and biological; psychological, cognitive and emotional; behavioural attributes and skills; social and cultural; physical environmental; and PA characteristics. Many of the variables listed here were assessed in the current study as well—table 1 presents an overview of the demographic and anthropometric variables, PA levels and readiness and health status (see also online supplemental table S1 for details regarding reported present diseases). Family incomes were binarised whether participant’s income was greater than the average family income in Japan (5 million yen). Education levels were also dichotomised to reflect whether participants graduated a university. These dichotomisations were used for interpretability and comparability as other countries may have different income levels and education systems (eg, middle school, secondary school, gymnasium). Dichotomisation leads to loss of information, but we did not expect substantial influences on the conclusions as the associations with PA were not explicitly hypothesised. There were several exceptions and deviations from Trost et al.12 First, for psychological characteristics, we measured variables related to personality, motivation (regulatory focus; hedonic and eudaimonic motives for activities in general), self-control and perceived social support (table 2). We placed social support under the psychological characteristic section as this was the only factor concerning social and cultural characteristics assessed in this study.

Other psychological variables were grouped under the theory-based category together with behavioural variables (table 2). These variables are typically drawn from the Transtheoretical model and Self-determination theory, encompassing motivation (specifically for exercise), self-efficacy, process of change, and decisional balance, which often tap into both psychological (cognitive) and behavioural aspects. As we wanted to distinguish between general psychological traits (eg, personality, motivation in general) and the constructs specifically developed in the context of PA and exercise, we decided not to merge the general psychological and theory-based characteristics.

Like Trost et al,12 we assessed physical environmental factors (table 3), representing the presence of sidewalks and bike paths as well as safety from crime and traffic. We expected that access to digital (not only physical) resources would also be a key factor to explain PA, and thus, the use of mobile health technology (apps and wearable activity trackers supporting PA) and technology acceptance were added to the list (table 3).

Statistical analysis

To find meaningful associations between the assessed variables, we estimated a network using the qgraph package (V.1.9.2)18 for R (V.4.2.2: R Core Team).19 This analysis estimated a graphical Gaussian model, in which an edge represents a partial correlation coefficient, indicating an association between a given pair of variables (nodes) after controlling for the other variables present in the network. Meaningful edges were selected by the graphical least absolute shrinkage and selection operator20 (GLASSO). The best-fit network structure was searched using the Extended Bayesian Information Criteria21 with the hyperparameter γ (a penalty term of the number of edges) set as 0.5.14 Prior to network estimation, a non-paranormal transformation22 was applied because the normality assumption was violated for certain (eg, binary) variables (see the online supplemental materials for a sensitivity analysis without the transformation, online supplemental figure S1). Each network diagram was depicted using the Fruchterman-Reingold algorithm,23 which determines edge lengths depending on the absolute values of the edge weights (ie, nodes with a higher partial correlation have a shorter edge length). On the estimated network, we first focused on the variables that had a direct association with PA. Second, we interpreted the indirect (two-step) associations leading to PA. Third, we computed two centrality indices: strength and expected influence, to explore how and which nodes were most closely associated with other nodes in the network. Strength centrality is given by the sum of absolute edge weights that a node has. Expected influence is the sum of signed weights, which considers the directions—positive and negative—of associations of a node.24 Both indices therefore quantify the importance of a given node in a network (ie, how many and how strong connections each node has). Edge accuracy (bootstrapped confidence intervals with 2.5 and 97.5% quantiles) and stability of the centrality indices25 were assessed using the bootnet package (V.1.5; see the online supplemental materials for the technical details, online supplemental figures S2 and S3).

Demographics

Demographic information is presented in table 1. The majority of participants were married (64%), had at least one child (62%), held a job (60%) and had obtained higher education (university or above; 43%). The mean body mass index score was 22.15 (SD=3.71) and the total PA level was 35.06 METs-hours/week (SD=56.94). As 9500 participants (50%) met the levels of PA recommended by the national guideline (ie, 23 METs-hours/week for adults aged <65 years; 10 METs-hours/week for older),26 our sample might be more active than the general population.

Direct associations with PA

The estimated network is illustrated in figure 1A, presenting all 295 edges (30%: 295 out of 990 possible edges) selected by the GLASSO algorithm. The mean absolute edge weight was 0.09. Panel B specifically illustrates the edges directly associated with PA (one-step neighbourhood). Of the 44 PA-correlates submitted to the network analyses, nine showed positive, direct associations: job/employment, edge weight=0.049, IQR = (0.036 to 0.062); stage of change, edge weight=0.256, IQR = (0.242 to 0.270); social support, edge weight=0.073, IQR = (0.059 to 0.087); self-efficacy, edge weight=0.062, IQR = (0.048 to 0.077); self-reevaluation, edge weight=0.024, IQR = (0 to 0.037); counter conditioning, edge weight=0.067, IQR = (0.055 to 0.081); intrinsic motivation for exercise, edge weight=0.037, IQR = (0.025 to 0.050); environment, edge weight=0.072, IQR = (0.057 to 0.083); and mHealth technology use, edge weight=0.092, IQR = (0.078 to 0.104). Two nodes showed negative, direct associations with PA: helping relationships, edge weight=−0.049, IQR = (−0.061 to –0.035); and external regulation for exercise, edge weight=−0.058, IQR = (−0.070 to –0.043). These results suggest that the variables that are closely associated with PA can be found at different (from intra-individual to environmental) levels: demographics (job/employment status); psychological characteristics and behaviour change strategies (social support, self-efficacy, self-reevaluation, counterconditioning, intrinsic motivation); physical environments; and technology (mHealth) uses. Not surprisingly, most of these variables are contextualised in PA and exercise (eg, intrinsic motivation for PA, but not motivation in general), and none of the general psychological traits (eg, Big-five personality, well-being) was directly associated with PA.

Request permission

Indirect associations with PA

Panel C in figure 1 represents the indirect associations, namely the edges related to the nodes that are directly associated with PA. This two-step neighbourhood network connected 40 nodes out of the 45 nodes, suggesting that most of the known PA-correlates are associated with PA at least indirectly. Among the nodes directly connected to PA, social support was found to be the most important hub for the other (indirectly associated) nodes, showing the largest number of edges in the two-step neighbourhood network. Social support bridged the associations with PA for demographics (age, gender and marital status); personality and general psychological characteristics (extraversion, agreeableness and pleasure motivation); theory-driven variables for behaviour change (pros and cons; integrated, identified, introjected and external regulation; and amotivation); and physical environments (number of vehicles).

Centrality

We found the highest strength centrality for age, identified regulation and self-reevaluation (figure 2), implying that these variables had the greatest absolute associations (aggregated) with the other nodes in the network (see the online supplemental materials for individual centrality scores). The highest expected influence (which considers the direction of edges) was identified for integrated regulation, self-reevaluation and promotion focus regulation, showing the largest positive associations with the other nodes in the sum. Neuroticism exhibited the most prominent negative expected influence, which suggests that this personality dimension is typically negatively associated with the variables in the network.

Request permission

Direct edges

The results identified nine variables, across individual, social and environmental levels, that are directly positively associated with PA: job/employment status, self-efficacy, perceived social support, intrinsic motivation, stage of change, counter conditioning, self-reevaluation, environment and mHealth technology use. As the edges reflect unique associations with PA after controlling for the other variables in the network, these nodes could be interpreted as proximate factors that are closely associated with PA. Four of the variables (self-efficacy, stage of change, counterconditioning and self-reevaluation) are from the transtheoretical model for PA,29 30 whereas intrinsic motivation and social support (as well as self-efficacy or competence) are highlighted in the social-cognitive theory31 and self-determination theory.32 Self-efficacy (defined as confidence in maintaining PA even in the presence of barriers) is known to increase as the stage of change progresses,33 from the precontemplation (having no intention to start exercise) to maintenance stage (having acquired an exercise habit). Counterconditioning and self-reevaluation are counted as the process-of-change strategies that help people adapt their behaviour to progress through the stages of change. These behavioural strategies focus on internal and external controls (eg, self-image; reward and punishment), which are relevant to improving self-regulation and autonomy as well as self-efficacy. Social support (eg, modelling by family and friends; support from an exercising partner) is another basis of self-efficacy and self-regulation,34 which are also known to be related to intrinsic motivation.35 The associations with physical environments and information technology (mHealth) uses are also in line with the literature.36 37 Our findings suggest that the PA-proximate factors can be found at different levels, but this does not mean that these factors work independently— instead, as the theories assumed, they are mutually associated with each other (see figure 1, Panel A), and there could be a cycle to enhance and maintain PA levels (eg, the use of behaviour change strategies supported by mHealth tools may lead to increased self-efficacy).

On the other hand, direct negative associations were identified for external regulation and helping relationships. The identified negative associations were intuitively difficult to interpret as both external regulation (eg, people around me (family, friends, doctors, etc) say I should take up exercise) and helping relationships (eg, your friends encourage you to exercise) conceptually overlap with perceived social support (eg, do you have a significant other, such as a family member, spouse, friend or colleague, who gives you advice or guidance on how to exercise?). External regulation and helping relationships almost exclusively relate to exercise recommendations from others, whereas social support is operationalised as various strategy supports that one receives, such as how-to tips, coexercising and appraisals. Consistent with our findings, a systematic review concluded that social pressure, like external regulation, tends to have little or even adverse effect on PA.35 This may suggest that the type and quality of support is significant to determine the direction of the association with PA.

Indirect edges

Another key finding is that the two-step neighbourhood network connected almost all the variables submitted to the network analyses (40 out of 45 nodes). Here, we included as many variables as possible, following the lists of empirically and theoretically relevant correlates of PA (eg, Trost et al).12 This implies that most of the known PA-correlates are at least indirectly (if not directly) associated with PA, which reassures us that each characteristic from individual to environmental levels (eg, demographics,4 motivation,5 self-efficacy,6 attitudes,7 and personality,8 environment38 and technology acceptance39) is significant to draw a profile of someone with an (in)active lifestyle in the context of the multilevel framework.1 It is also noteworthy that social support worked as the most important hub for PA, connecting the largest number of nodes (including marital status, extraversion personality) in the two-step neighbourhood network. Inter-person factors, particularly social incentives (collaborations and competitions), are known to be effective in improving PA40 although large heterogeneity has been documented.41 Causal inference is not possible with our data, but our results may suggest that family status and pro-social psychological characteristics (eg, extraversion) are prerequisites for social support to function effectively. We see the current network analysis as a hypothesis-generation process, and the identified in-direct associations will serve as new hypotheses to be tested in future research using a more hypothesis-driven approach (eg, structural equation modelling) with an experimental or longitudinal study design for establishing temporal and causal relationships.

Centrality indices

Furthermore, the centrality analyses showed the highest strength for age, which was most closely associated with other variables in the network (followed by identified regulation and self-reevaluation). Previous studies have already identified similar associations between age and PA-correlates: for example, personality,42 exercise motivation,43 barriers44 45 and mobile device use.46 Conversely, environmental barriers have been documented commonly across different age groups.45 Interestingly, age did not have a direct edge to PA in the estimated network—individual and social characteristics may differ across age groups, but age per se works as a pure indirect predictor, not directly informing PA levels after controlling for other PA-correlates. The highest expected influence was found for integrated regulation, followed by self-reevaluation and promotion focus regulation, which had mostly positive associations with other nodes in the network. On the other hand, neuroticism showed the greatest negative expected influence. These findings imply that motivation and process of change may play—as the transtheoretical model and self-determination theory assumed—key roles in active lifestyles, whereas neuroticism and mental health issues could be a barrier in maintaining individual and environmental factors supporting PA—indeed, neuroticism is related to exercise barriers.47

Limitations

Several significant limitations should be noted when interpreting our results. First, the cross-sectional nature of the study limits the ability for causal inference. Bauman et al1 emphasised the significance of identifying determinants, beyond correlates, to develop and establish an intervention that helps efficiently increase engagements in PA. Although we believe that our findings refined the list of PA-related factors by identifying the direct (and indirect) associations, a longitudinal study is warranted to explore potential prospective (and causal) associations among the correlates. Second, we aimed to analyse PA-related variables as comprehensively as possible, but exhaustive assessments with any relevant measures were not possible for pragmatic reasons. Critically, the network structure may vary depending on the variables submitted to the analysis, which calls for careful replication with different selections of variables. Third, we solely relied on self-report measures, which may be affected by reporting bias. It would be significant to apply objective assessment methods (eg, accelerometers) in future replications.